Transition State Theory Methods To Measure Diffusion in Flexible

Dec 18, 2015 - Transition state theory (TST) methods are useful for predicting adsorbate diffusivities in nanoporous materials at time scales inaccess...
0 downloads 0 Views 1MB Size
Subscriber access provided by CMU Libraries - http://library.cmich.edu

Article

Transition State Theory Methods to Measure Diffusion in Flexible Nanoporous Materials: Application to a Porous Organic Cage Crystal Jeffrey Camp, and David S. Sholl J. Phys. Chem. C, Just Accepted Manuscript • DOI: 10.1021/acs.jpcc.5b11111 • Publication Date (Web): 18 Dec 2015 Downloaded from http://pubs.acs.org on December 19, 2015

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

The Journal of Physical Chemistry C is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Transition State Theory Methods to Measure Diffusion in Flexible Nanoporous Materials: Application to a Porous Organic Cage Crystal

Jeffrey S. Camp and David S. Sholl* School of Chemical & Biomolecular Engineering Georgia Institute of Technology, Atlanta, GA 30332-0100, USA

*

Corresponding author. email: [email protected] 1 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 37

Abstract Transition state theory (TST) methods are useful for predicting adsorbate diffusivities in nanoporous materials at timescales inaccessible to molecular dynamics (MD). Most TST applications treat the nanoporous framework as rigid, which is inaccurate in highly flexible materials or where adsorbate dimensions are comparable to the size of pore aperatures. In this study, we demonstrate two computationally efficient TST methods for simulating adsorbate diffusion in nanoporous materials where framework flexibility has a significant influence on diffusion. These methods are applied to light gas diffusion in porous organic cage crystal 3 (CC3), a highly flexible molecular crystal that has shown promise in gas separation applications. Diffusion in CC3 is modeled as a series of uncorrelated adsorbate hops between cage molecules and the voids between adjacent cages. The first method we applied to compute adsorbate hopping rates in CC3 is implicit ligand sampling (ILS). In ILS TST, hopping rates are calculated in an ensemble of rigid framework snapshots captured from a fully flexible MD trajectory of the empty CC3 structure. The second TST method we applied is umbrella sampling (US), where hopping rates are computed from a series of biased MD simulations. Our ILS and US TST calculations are shown to agree well with direct MD simulation of adsorbate diffusion in CC3. We anticipate that the efficient TST methods detailed here will be broadly applicable other classes of flexible nanoporous materials such as metal–organic frameworks. Keywords: transition state theory; hopping rate; diffusion; molecular dynamics; porous organic cage; cage crystal 3; framework flexibility

2 ACS Paragon Plus Environment

Page 3 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

1. Introduction Adsorbates within nanoporous materials such as zeolites and metal–organic frameworks experience strong confinement effects that can result in transport properties dramatically different than bulk phases. Atomistic simulations of these adsorbed phases can provide insight on the kinetics of transport in nanopores where experimental characterization would be challenging.1 For example, molecular dynamics (MD) simulations can give adsorbate diffusivities, which are an important predictor of material performance in both equilibrium and kinetic separation applications.2 These MD simulations frequently treat the crystalline framework as a rigid body.3-4 This approach greatly reduces the computational cost of MD because adsorbate–framework potential energies can be mapped to a grid and do not need to be recomputed over the course of a simulation. The rigid framework assumption may be reasonable for stiff frameworks, or where adsorbate dimensions are significantly smaller than pore apertures.5 In highly flexible materials or where adsorbates are comparable in size to pore apertures, framework flexibility has a significant influence on diffusion.6 One such class of flexible materials fulfilling both these criteria is porous organic cages (POCs), which crystallize in the solid state without forming intermolecular covalent or coordination bonds.7-8 The weak nature of these intermolecular forces make POCs inherently flexible.9 Flexibility increases the cost of MD simulations of POCs and other flexible nanoporous materials because many framework degrees of freedom must be updated with each MD step. One of the first POCs synthesized, known as “cage crystal 3” (CC3),10 has a number of desirable properties that make the material potentially useful in kinetic separation applications. CC3 forms a 3-D diamondoid pore network with a BET surface area11 of 409 m2 g-1 which is thermally 3 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 37

stable to 398 C10, stable in boiling water12, and reversibly adsorbs over 20 wt% water.12 Gas phase MD simulations of adsorbates within isolated cage molecules were used to predict that mesitylene molecules cannot overcome a reorientational barrier to diffusion in CC3 while smaller aromatics such as 4-ethyltoluene can. These predictions were consistent with single phase gas uptake experiments which demonstrated that CC3 can adsorb a significant amount of 4-ethyltoluene, but no mesitylene.13 Subsequent MD studies of diffusion in crystalline CC3 at 300 K showed that 4-ethyltoluene molecules hop between cages and inter-cage voids several times over the course of a 20 nanosecond simulation, but mesitylene molecules are completely immobile. Recently, CC3 was reported to have attractive performance in the adsorptive separation of Kr, Xe, and Rn from air.14 Although qualitative insights from MD can be valuable, typically hundreds of adsorbate hops must be observed in MD to yield a well converged diffusion coefficient. This “MD timescale problem” is a well known challenge in the simulation of diffusion in nanoporous materials.15 In CC3, this challenge is also relevant for smaller spherical adsorbate molecules. Evans et al. performed MD simulations of light gas diffusion to evaluate the performance of mixed-matrix membranes with CC3 and other porous organic cage additives.16 The diffusion coefficient of CH4 in CC3 measured in that study at 298 K and 10 bar adsorbate loading pressure was 2.16×106

cm2s-1, which corresponds to only a few CH4 hops per nanosecond.16 Similarly, Holden et al.

found that the diffusion coefficient of Xe in CC3 at 298 K is 1.83×10-6 cm2s-1. It is computationally expensive to measure diffusion coefficients on this order of magnitude using MD, particularly at low adsorbate loadings. Larger adsorbates of interest in CC3 such as 1phenylethanol14 are expected to diffuse orders of magnitude slower, making direct MD simulation of diffusion infeasible. 4 ACS Paragon Plus Environment

Page 5 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

The primary objective of this paper is to demonstrate two numerically efficient transition state theory (TST) methods used to compute the self-diffusivities of light gases in CC3 in regimes of low temperature and infinite dilution where diffusion is inaccessible to straightforward MD. These methods are an alternative to more complicated transition path sampling methods used to simulate slow diffusion in flexible nanoporous materials.17 We also describe techniques to analyze the relationship between pore aperature (window) sizes, adsorbate hopping activation energies, and adsorbate hopping rates in CC3. We anticipate that these methods will be broadly applicable to a wide variety of adsorbates in porous molecular crystals and other classes of flexible nanoporous materials. 2. Modeling diffusion in CC3 2.1. Structure and force fields A 2×2×2 supercell of the homochiral CC3-R structure (Cambridge Structural Database: PUDXES)18 was used throughout this work. All window sizes reported here were calculated by the arene carbon method described by Chen et al.14 for consistency with window sizes previously reported in CC3.14, 19 This approach is convenient, but less rigorous than a definition of window size based on percolating sphere diameters20 or Voronoi decomposition.21 Cage adsorption force field (CAFF)14: The cage-adsorption force field (CAFF)14 was used to model adsorbate-cage interactions for CH4 and the noble gases. CAFF is based upon DREIDING22 with rescaled epsilon parameters to fit adsorption experimental data.14 United atom models of carbon disulfide (CS2) and sulfur hexafluoride (SF6) were represented using experimentally derived 12-6 Lennard Jones parameters from the literature.23 Parameters from EPM224 and UFF25 for CO2 and the framework, respectively, were combined with the Lorentz– Berthelot mixing rules to model carbon dioxide–framework dispersion interactions. CSFF partial

5 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 37

charges were used to model electrostatic interactions with CO2. CAFF parameters were also used to model adsorbate–adsorbate interactions. All adsorbate-cage and adsorbate-adsorbate interactions were truncated at 10 Å. Cage-specific force field (CSFF)26: CSFF was parameterized by Holden et al.26 to describe the framework dynamics of CC3 and other porous organic cages. In CSFF, intramolecular nonbonded interactions are completely excluded, meaning that the 168 atoms within individual cage molecules do not interact by Lennard-Jones or Coulombic forces. To describe intermolecular forces between individual cage molecules, CSFF uses partial charges and scaled 9-6 Lennard Jones parameters from the polymer consistent force field (PCFF).27 Following the original CSFF report, the dispersive portion of Lennard-Jones potential was scaled by a factor of 1.20. All CSFF intermolecular Lennard-Jones interactions were truncated at 10 Å. Coulombic interactions were computed pairwise to 10 Å and a long range particle-particle mesh Ewald

correction was used thereafter. Bonded forces were applied as described in the original CSFF report with the following exceptions: two angle and three torsion parameters absent from the original CSFF report26 were adapted directly from PCFF by Holden and coworkers and used throughout this work. These parameters are necessary to reproduce the geometry of the cyclohexyl groups on the CC3 molecules. The coefficient leading the class2 trigonometric dihedral potential was also modified, as described in the Supporting Information. 2.2. Energy minimization of the CC3 structure The atomic positions of the CC3 structure were minimized with the CSFF force field using a damped dynamics algorithm28 to a tolerance of 10-11 kcal mol-1. Independent minimization runs were performed at variable lattice constants created by isotropically scaling the CC3 atomic

6 ACS Paragon Plus Environment

Page 7 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

positions by factors ranging 0.97 to 1.03 (corresponding to changes in unit cell volumes between -9% and +9%).

Figure 1. Energy minimized CC3 window sizes and potential energies as a function of unit cell volume scaling.

Figure 1 shows the energy minimized window sizes and total potential energies for unit cell volumes ranging from 9% smaller to 9% larger than the experimental CC3 structure. At the experimentally refined lattice constants (0% volume scaling), the CC3 windows relax to 3.80 Å, about 5% larger than the window size of the experimental structure (3.62 Å). The potential energy of the relaxed CC3 structure at the experimental unit cell volume is slightly higher (0.05%) than the relaxed CC3 structure found after applying a 1.5% volumetric contraction in the unit cell. Over the range of unit cell volumes considered small differences in potential energy (less than 1.5%) are associated with significant changes in the relaxed CC3 window size (3.49 – 3.97 Å). In molecular dynamics simulations using the flexible CSFF force field, frameworks, the CC3 windows assume many different conformations as the simulation progresses.

7 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 37

Figure 2. Normalized histogram of CC3 window sizes in fully flexible NVT MD at 300 K. The experimental window sizes (3.62 Å) is indicated with an arrow.

After 10 ps of NPT equilibration at 300 K, the CC3 structure contracts by 4.2%. Figure 2 shows the normalized window size distribution at 300 K in the empty CC3 structure over a 100 ps NVT production period at these NPT equilibrated lattice constants. The window size distribution observed in NPT simulations is essentially identical to the NVT results shown in Fig. 2 (data not shown). The window size distribution is very well fit by a normal distribution, which is a common feature of many nanoporous materials.6, 16 The window size of the energy minimized structure at the corresponding lattice parameters (3.61 Å at 4.2% contraction) is close to the mean

of the distribution (µ = 3.59 Å). The fit parameters are in reasonably good agreement with the

CC3 window size distribution reported in the original CSFF publication, which is centered near µ = 3.75 Å.26 2.3. Diffusion coefficients by MD The diffusion coefficients of Kr, CH4, and CO2 were measured with NVT molecular dynamics at 300 K using a Nose-Hoover thermostat and a 0.5 fs timestep. For the other adsorbates and temperatures considered below, diffusion was too slow to be readily simulated with straightforward MD. Each of the adsorbates simulated with MD was loaded to a concentration of 2 adsorbate molecules per unit cell, which each contain 8 CC3 molecules. This corresponds to a 8 ACS Paragon Plus Environment

Page 9 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

total of 16 adsorbate molecules in a 2×2×2 CC3 supercell, or a ratio of 1 adsorbate atom for every 4 porous organic cage molecules. The experimentally measured saturation loading of CC3 for Kr and Xe is 2.1 and 2.69 adsorbate atoms per cage, respectively, at room temperature.14 We expect that our simulated loading of 0.25 adsorbates per cage results in negligible adsorbate– adsorbate interaction effects on self diffusivities measured by MD. This facilitates comparison to TST calculations performed at infinite dilution. The loaded structure was subject to a 20 ps NPT equilibration period at a pressure of 1 bar followed by a 20 ps NVT equilibration at the equilibrated cell volume. Twenty eight (28) independent 1 nanosecond NVT simulations were averaged and then fit to the Einstein relation to give the self diffusivity, Ds. In each case, the MSD vs. time curve was renormalized to zero by subtracting 4 Å2 from the MSD, which is associated with the ballistic movement of adsorbate atoms within individual cage molecules. The mean–squared displacements plots and the associated fits are shown in Supplementary Figures 2-4. 2.4. Calculating diffusion coefficients from TST hopping rates Diffusion in CC3 proceeds by activated adsorbate hopping of molecules between a cage and the void space separating adjacent cages, also known as “window cavities”.14 The self-diffusion coefficient is given by weighting the hopping rates out of the 4-coordinated cage sites (kC→V) and the 2-coordinated void sites (kV→C) by the equilibrium probability of occupying cages or voids (PC and PV, respectively)29:  =

1

4→  + 2→  ) 6

1)

Here, λ is the hopping distance between cages and voids, which is 5.32 Å in the NPT equilibrated structure at 300 K. Since the likelihood of occupying a cage or void is directly proportional to 9 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 37

the residence time in each site (the inverse of the hopping rate), the occupancy probabilities can be eliminated from eq. 1, giving:  =

1 → 2→

4   → + 2    6 2→ + → 2→ + → →

2)

The hopping rates kC→V and kV→C were calculated using the 1-dimensional TST methods described below. 3. Computing hopping rates using TST 3.1. TST in the rigid experimental structure The local framework flexibility associated with window size deformation is understood to strongly influence diffusion in CC3.16, 30-31 We first performed simulations of spherical adsorbate diffusion in the rigid experimental structure to construct an appropriate reaction coordinate and to establish a basis of comparison for our fully flexible simulations. The 1-dimensional reaction coordinate (denoted q) used in all of our simulations is coincident with the vector that connects the centers of mass of two adjacent cage molecules. The coordinate system along this line is centered (q = 0) at the midpoint between these positions in the void space that spans adjacent cages. Along this line, the centers of mass are located at q = ± 5.4 Å in the experimental structure and each window is centered at q = ± 2.5 Å. For convenience, we constructed a reaction

coordinate between two cage molecules that do not span periodic boundaries in the 2×2×2 CC3 supercell. To determine the free energy profile, F(q), for adsorbate hopping along this reaction coordinate, we calculated the mean energy of insertion of spherical adsorbate molecules using CAFF parameters in planes orthogonal to the reaction coordinate:  ) = −  ln〈

!∆# 〉

%

(3) 10

ACS Paragon Plus Environment

Page 11 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Here, the brackets denote averaging the Boltzmann factor over square grids of 0.2 Å resolution perpendicular to the reaction coordinate at positions along q. Each of these squares is 11.4 × 11.4 Å across (57 × 57 insertions) to fully enclose the cage cavities. These square planes were

positioned at 0.2 Å intervals to form a contiguous 3-dimensional rectangular cuboid grid oriented parallel to q. Figure 3 illustrates the construction of the square grid “slices” orthogonal to the reaction coordinate.

Figure 3. The TST reaction cordinate superimposed on the experimental CC3 structure. Every 15th square grid slice is shown.

The cage to void hopping rate kC→V was calculated by applying 1-dimensional TST15: →

   !* % ) = &' 2() ,  !* %) - ./01 ∗

4)

Here, m is the mass of the adsorbate molecule and κ is the Bennett-Chandler dynamic correction factor. One dimensional TST has been applied extensively to model diffusion in rigid zeolites and metal organic frameworks.32-33 We have assumed a dynamic correction factor of 1 throughout this work, which was shown to be a good approximation for spherical adsorbates in LTL and LTA-type zeolites.15 F(q*) is the free energy at the transition states, which correspond to the two local maxima in F(q). The denominator of eq. 4 was evaluated by integrating over the 11 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 37

points on the reaction coordinate associated with the respective cage microstates. F(q) is symmetric in the rigid experimental structure (Figure 4b), meaning that kcage 1→void and kcage 2→void are equal. We evaluated the void to cage hopping rate kV→C by integrating over the void microstate, corresponding to the points on F(q) between q1* and q2*. To visualize the cage and void microstates in CC3, we applied the Fernand–Meyer watershed segmentation algorithm34 to the 3-dimensional adsorbate potential energy grid described above. In 2D image processing, this algorithm is used to segment images into regions based on pixel intensity. For example, a 2D grayscale image can be considered a 3D topographic map where altitude at each point is proportional to pixel intensity. The Fernand-Meyer algorithm identifies the 2D “catchment basins” and 1D “ridge lines” that separate adjacent regions in this topographic projection. In our 3-dimensional application, the adsorbate potential energy grid in each voxel is mapped to a grayscale pixel intensity. The Fernand-Meyer algorithm then identifies the 3D microstates and 2D dividing surfaces in potential energy space. This procedure is similar to the Henkelman et al. algorithm for grid based Bader decomposition of charge density.35

12 ACS Paragon Plus Environment

Page 13 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Figure 4 (a) Microstates in the rigid experimental CC3 structure revealed by the watershed segmentation algorithm. The 4-coordinated cage sites are shown in blue while the 2-coordinated void sites are shown in red. (b) 1-dimensional free energy profile for Xe hopping at 300 K in the rigid experimental structure.

Figure 4(a) shows the result of watershed segmentation of the Xe potential energy grid between adjacent CC3 cages calculated with eq. 3 for the experimental CC3 structure. The dividing surfaces that separate the cage microstates (shown in blue) and the void microstates (shown in red) are associated with the transition state free energies F(q*) on the corresponding 1dimensional Xe free energy profile shown in Figure 4(b). In the rigid experimental structure, the free energy profile, F(q), and associated hopping rates for a given adsorbate can be calculated at arbitrary temperatures by changing T in eq. 3 and eq. 4. Between any two temperatures T1 and T2, the cage to void hopping rates kV→C can be fit to the Arrhenius equation to yield to activation energy for cage to void hops EA,C→V: 13 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

23,→ = −5

ln6→ 7 )8 − ln6→  )8 7

7

− 

7

Page 14 of 37

5)

The Arrhenius prefactor is computed in a similar way. The same procedure is used to determine the activation energy for void to cage hops EA,V→C. 3.2 Flexible implicit ligand sampling TST To calculate adsorbate hopping rates in the fully flexible CC3 structure, we first applied a variation of the implicit ligand sampling method36 developed to study diffusion in flexible nanoporous materials.6 In this approach, hopping rates are determined by TST in an ensemble of rigid frameworks (“snapshots”) captured from a fully flexible MD trajectory of the empty porous material. The overall flexible hopping rates are then calculated by averaging the hopping rates calculated in each individual framework conformation. Awati et al. have shown that this approach is quantitatively accurate for diffusion of light gases in small pore zeolites.37 This method assumes that adsorbates do not significantly influence the motion of the crystalline framework. Snapshots of the fully flexible CC3 structure were generated by running 20 ps of NPT equilibration at 1 bar followed by a 100 ps NVT production period. Snapshots were saved at 100 fs intervals, producing ensembles of 1,000 snapshots at temperatures of 200, 250, and 300 K. The NPT equilibration period results in contraction of the CC3 framework by 5.3% at 200 K, 4.9% at 250 K, and 4.2% at 400 K relative to the experimental structure. The brief molecular dynamics production period used here means that the ensemble of snapshots will capture only local flexible framework vibrations, such as changes in CC3 window sizes, not possible phase transitions that occur at longer timescales. CC3 is not known to undergo any large changes in the structure of its unit cell, but deformations of this type could pose a challenge to applying implicit

14 ACS Paragon Plus Environment

Page 15 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

ligand sampling to materials such as cage crystal 1 which can switch from a porous to a nonporous phase.38 In each flexible snapshot, we calculated the free energy profile, F(q), by the same grid method as described above for the rigid experimental structure. The hopping rates in each snapshot were determined by 1-dimensional TST using eq. 4. Within each snapshot, we calculated two independent cage to void hopping rates (kcage 1→void and kcage 2→void ) and two independent void to cage hopping rates (kvoid→cage 1 and kvoid→ cage 2 ) in each of 16 cage-void-cage pairs that do not cross periodic boundaries. This yielded samples of 32,000 kC→V and 32,000 kV→C hopping rates when performed over 1,000 snapshots. Independent rate samples must be calculated for each adsorbate at each temperature.

Figure 5 (a) Microstates revealed by the FernandMeyer algorithm in a snapshot from flexible NVT dynamics at 300 K. (b) Corresponding 1-dimensional free energy profile for Xe hopping.

15 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 37

Figure 5(a) depicts the cage and void microstates in a snapshot collected from NVT dynamics at 300 K using CSFF. The void microstate spanning cage 1 on the left and cage 2 on the right is deformed with respect to the symmetrical void microstate in the rigid experimental structure shown in Figure 3(a). The cage 1 window coincident with the dividing surface between cage 1 and the void is 3.08 Å in diameter in this snapshot. This leads to a wide dividing surface between cage 1 and the void relative to the CSFF experimental structure, which has windows of 3.62 Å in diameter. In the corresponding 1-dimensional free energy profile in Figure 5(b), the free energy barrier for Xe hops from cage 1 is 74 kJ mol-1, compared to 23 kJ mol-1 in the experimental structure. In contrast, the window between cage 2 and the void is 3.92 Å in diameter, leading to a lower free energy barrier (13 kJ mol-1) for hops between cage 2 and the void. The difference in free energy barriers gives a kcage 1→void hopping rate 10 orders of magnitude slower than kcage 2→void

in the CC3 framework conformation represented by this snapshot. This enormous

difference in rates hints at the important role of framework flexibility in determining the overall diffusivity of species like Xe in CC3. The overall flexible hopping rates kC→V and kV→C can be estimated from the sample average of the 32,000 independent rates collected for each adsorbate and temperature condition. This approach can lead to numerical uncertainty, however, when a few exceptionally high rate observations have a dominant influence on the overall hopping rate. For example, in our sample of SF6 kC→V rates collected at 200 K, the 2 highest observations (out of 32,000) contributed over 80% to the numerator of the sample average. This high sample variance leads to unacceptably large numerical uncertainty in the sample average. We addressed this problem by inferring the structure of the distribution of hopping rates from each hopping rate sample as described below.

16 ACS Paragon Plus Environment

Page 17 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Histograms of the various kC→V and kV→C hopping rate samples collected were not related to an obvious statistical distribution. We investigated further by calculating the activation energies, EA, underlying each hopping rate in our kC→V and kV→C samples with eq. 5 in the same way as in the rigid experimental structure. In this case, T1 is the physically meaningful temperature associated with the NVT snapshot collection while T2 is an arbitrary second reference temperature. The Arrhenius prefactor, A, underlying each rate was computed in a similar way. For example, in the snapshot shown in Figure 5, the Xe kcage 2→void hopping rate is 2.51×109 s-1 at 300 K, which can be decomposed into a hopping activation energy of 8.28 kJ mol-1 and an Arrhenius prefactor of 6.94×1010 s-1. This activation energy is significantly lower than in the experimental structure (17.3 kJ mol-1) because of the favorable window geometry between cage 2 and the void in this snapshot. The resulting histograms of 32,000 hopping activation energies underlying each sample of 32,000 kC→V and kV→C rates are well described by log-normal distributions. Figure 6 shows the distribution of Xe kC→V hopping activation energies at 300 K and the associated fit to a lognormal distribution computed with the MATLAB maximum likelihood estimation function.39 Similar distributions were fit to each of the kC→V and kV→C rate samples collected for each adsorbate at different temperature conditions. The log-normal location parameter µ and scale parameter σ for each rate sample with associated 95% confidence intervals are given in Supplementary Tables S7-S18.

17 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 37

Figure 6. Distribution in activation energies for krypton cage to void hops at 300 K with associated fit to a log-normal distribution. Quantities in parentheses indicate 95% confidence intervals.

With the parameters µ and σ for a given rate sample, the expectation value of the overall fully flexible hopping rate kC→V or kV→C was found by integrating the Arrhenius rate equation over the respective log-normal probability density function: H

 = :̅ < I

1

− ln = − D) −= exp C E exp F G -= 2? 5 =√2(?

6)

Here, :̅ is the Arrhenius prefactor averaged over the 32,000 individual Arrhenius prefactors calculated for each rate sample. This method neglects the effect of covariance between A and the hopping activation energy, which was found to be small (Figure S2). Uncertainties in the overall flexible hopping rates were estimated by recomputing kC→V or kV→C at the low and high values of µ and σ from the 95% confidence intervals. When this fitting procedure is applied to fast moving adsorbates such as Kr, the overall flexible hopping rate from eq. 6 (kC→V = 6.28×109 ± 0.04×109 s-1 at 300 K) is close to the sample average of 32,000 kC→V rates (7.33×109 ± 0.25×109 s-1) . For larger adsorbates, flexible hopping rates from eq. 6 have much lower numerical uncertainty than given by simple sample averages. For example, the activation energies for SF6 cage to void hopping at 300 K were fit to a log-normal 18 ACS Paragon Plus Environment

Page 19 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

distribution with µ = 11.89, σ = 0.4742, A = 2.12×1011 s-1. Here, µ and σ are normalized to a base unit of 1 J mol-1 such that eµ yields the most probable diffusion activation energy in units of J mol-1. The expectation value of the kC→V hopping rate for this activation energy distribution given by eq. 6 is 6500 s-1. The population variance σ2k of the rate distribution calculated from the cumulative distribution function of kC→V is 1.662 × 1012 s-2 (Supplementary equations S10 and S11) .40 By the central limit theorem, the sample averages of samples of 32,000 rates randomly drawn from this rate distribution would have a standard deviation of 7200 s-1. Therefore, the means of many such rate samples would be deviate from the expectation value by over an order of magnitude. Using our maximum likelihood estimates of the 95% confidence intervals on µ and σ, we calculate an uncertainty of 6500 ± 500 on the value of the SF6 cage to void hopping rate at 300 K. This is a significant improvement over sample averages. 3.3 Flexible umbrella sampling TST In the implicit ligand sampling method described above, the dynamical problem of rate constant calculations in the fully flexible CC3 structure is reduced to a series of static rate constant calculations in framework conformations sampled from fully flexible dynamics. This method makes the assumption that interactions between adsorbates and the CC3 framework are negligible. This precludes a cooperative diffusion mechanism in which the CC3 window dynamics are influenced by the presence of adsorbates. This assumption may be invalid for adsorbate molecules significantly larger than the size of the CC3 window in the experimental structure. To overcome this limitation of the implicit ligand sampling method, we must calculate a free energy profile, F(q), that takes into account adsorbate-framework interactions in the fully flexible CC3 structure. This free energy profile could in principle be calculated by histogramming the 19 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 20 of 37

position of an adsorbate over a very long fully flexible MD simulation to yield the adsorbate probability density function along the reaction coordinate P(q), which is directly proportional to F(q).41 However, since the adsorbate rarely visits the high energy regions near the transition state, this method would be no more computationally efficient than computing diffusivities with straightforward MD. Umbrella sampling is a technique to improve computational sampling of systems with high free energy barriers by means of (typically harmonic) restraint forces.42 At intervals along the reaction coordinate known as “umbrellas”, these spring forces constrain the motion of the adsorbate (or other component of interest) to a region near the spring center. In each umbrella, an independent molecular dynamics simulation is performed to histogram adsorbate positions under the influence of the spring. The unbiased probability distribution is then reproduced from the overlapping histograms from each window, yielding the free energy along q. Umbrella sampling has been used, for example, to study benzene diffusion in rigid4 and flexible43 zeolites. We implemented umbrella sampling using the LAMMPS collective variables library.44 To reconstruct the free energy profile F(q), we histogrammed the adsorbate position over 24 umbrellas distributed over evenly spaced intervals from q = -6.4 to q = 6.4 Å . The adsorbate

atom was confined to each umbrella by a 5 kcal mol-1 Å-2 harmonic restraint acting along q. In

each umbrella, the adsorbate atom was placed at the harmonic restraint center, velocities were initialized from the Maxwell-Boltzmann distribution, and a 10 ps fully flexible NVT equilibration was run. Following equilibration, the positions of the adsorbate along q within each umbrella was recorded at 0.5 fs intervals over a 100 ps NVT production period.

20 ACS Paragon Plus Environment

Page 21 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Figure 7(a) depicts the 24 histograms of Xe positions used to determine F(q) for Xe hopping at 300 K. The free energy profile reconstructed was reconstructed from these histograms with the Grossfield et al. Weighted Histogram Analysis Method (WHAM) code.45-46 The free energy profile for Xe hopping at 300 K is shown in Figure 7(b). At 300 K, the free energy barrier for Xe cage to void hops (16 kJ mol-1) is lower than in the experimental structure (23 kJ mol-1, shown in Figure 3(b)). This is not surprising; it reflects the observation that hopping is dominated by framework conformations where the hopping barrier is considerably lower than the barrier in the rigid structure. The overall fully flexible hopping rates kC→V and kV→C from umbrella sampling were calculated with the 1-dimensional TST (eq. 4).

Figure 7 (a) Frequency histograms (umbrellas) for a Kr atom confined to 24 windows along the reaction coordinate by harmonic potentials at 300 K. (b) Reconstructed Kr free energy curve F(q).

21 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 22 of 37

TST calculations on polyatomic adsorbates are expensive using the implicit ligand sampling method because many rotational orientations must be integrated over at each grid voxel. In umbrella sampling, these rotational degrees of freedom are efficiently sampled by molecular dynamics. We used umbrella sampling to determine the free energy profile F(q) for CO2 diffusion in CC3 (Figure 8). In this case, we found that CO2 has a qualitatively different diffusion mechanism than spherical adsorbates. The free energy maxima are not located at the windows, but significantly closer to the midpoint between the adjacent cages. A similar energy profile is seen for CO2 diffusion in 8-member ring zeolites.41 The shallow energy minima at the void results in a high void to cage hopping rate kV→C.

Figure 8. Free energy profile F(q) for CO2 diffusion at 300 K by umbrella sampling.

4. Results and discussion 4.1. Infinite dilution diffusion coefficients The three 1-dimensional TST methods we used make different simplifying physical assumptions that reduce computational costs relative to straightforward MD. The simplest method we applied is 1-dimensional TST in the rigid experimental structure. The most physically realistic method we used is umbrella sampling of the fully flexible structure, since this approach considers all degrees of freedom relevant for diffusing molecules. Table 1 shows Ds from the two fully 22 ACS Paragon Plus Environment

Page 23 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

flexible TST methods at 300 K and the ratio of the flexible umbrella sampling TST Ds to Ds in the rigid experimental structure. The smallest spherical adsorbates we considered, Kr and CH4, have flexible umbrella sampling (US) and ILS TST self diffusivities only a factor of about 2 higher in the flexible CC3 structure versus the rigid CC3 structure at 300 K. In the cases of Xe and Rn, consideration of framework flexibility results in over an order of magnitude faster diffusion than in the rigid experimental structure. CS2 and SF6 are essentially immobile in the CSFF rigid experimental structure, but diffuse readily in the fully flexible CC3 structure. These results show that considering framework flexibility can be crucial for capturing even the qualitative nature of adsorbate diffusion in CC3.

Table 1. Ds for adsorbates in CC3 at 300 K computed by umbrella sampling (US), implicit ligand sampling (ILS), and by TST in the rigid experimental structure. adsorbate (CAFF)

flexible US Ds at 300 K (cm2/s )

flexible ILS Ds at 300 K (cm2/s )

Kr CH4 Xe Rn CS2 SF6

1.81×10-5 5.08×10-5 2.87×10-6 1.12×10-6 6.02×10-8 3.49×10-10

1.49×10-5 3.20×10-5 1.38×10-6 4.03×10-7 1.47×10-8 2.53×10-11

ratio of flexible US Ds to rigid exp. Ds 1.7 2.3 26 170 1.6×106 1016

Straightforward MD is the most rigorous description of diffusion for a given classical force field because no mechanistic assumptions are made about adsorbate diffusion dynamics. To test the quantitative accuracy of ILS and US TST, we compared each method to MD at 300 K (Figure 9). For each adsorbate at 300 K shown in Figure 9, the numerical uncertainties in the MD and TST self-diffusivities are on the order of the symbol size.

23 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 24 of 37

Figure 9. Comparison between TST diffusion coefficients from umbrella sampling (filled symbols) and implicit ligand sampling (empty symbols) and straightforward MD at 300 K.

Umbrella sampling and implicit ligand sampling TST overpredict the diffusion coefficients of Kr, CH4, and CO2 by factors ranging from 1.2 to 2.4. This systematic overprediction of Ds makes sense because TST hopping rates uncorrected by a Bennett-Chandler transmission coefficient, κ, should typically be greater than or equal to the true value (eq. 4). The ILS TST measurements of DS are consistent with transmission coefficients (κ) of approximately 0.85 for the overall kC→V and kV→C flexible hopping rates. The higher US TST Ds values are consistent with transmission coefficients (κ) of approximately 0.6, which are reasonable for flexible simulations where the precise location of the dividing surface is expected to vary during the simulation. Overall, both TST methods give reasonable approximations of these fast moving adsorbates.

24 ACS Paragon Plus Environment

Page 25 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

Figure 10. (a) Self diffusion coefficients by umbrella sampling (b) Ratio of US TST DS to ILS TST DS.

Figure 10(a) shows Ds from US TST for each adsorbate as a function of inverse temperature. Each set of data in CC3 is well described by an Arrhenius equation. Figure 10(b) gives the ratio of Ds from US TST to ILS TST. For adsorbates other than CS2 and SF6, there is only an insignificant difference between the US TST and ILS TST measurements. In each case the US TST measurement is higher, but this could be attributable to a lower Bennett-Chandler correction in the US TST method. For CS2 and SF6, umbrella sampling TST gives self-diffusivities significantly higher than ILS TST across the temperatures considered. This discrepancy may be due to the ILS TST assumption that adsorbates do not influence the motion of the CC3 framework. This precludes a concerted mechanism in which CC3 windows expand in the presence of nearby adsorbate molecules, lowering the EA for adsorbate hopping. TST based on umbrella sampling does not make any assumptions about decoupling between adsorbate and framework motions. By comparing the US derived Ds (which accounts for adsorbate-framework 25 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 26 of 37

interactions) to the ILS derived Ds (which assumes framework motions are independent of adsorbates), we can quantify the importance of these interactions on diffusion. For SF6, this factor contributes to up to an order of magnitude difference between US TST and ILS TST derived self diffusivities. Adsorbate-framework interactions could be more significant for larger polyatomic adsorbates such as aromatics and in materials more flexible than CC3. 4.2. Influence of CC3 window size on diffusion Analysis of individual framework conformations from our ILS TST calculations (Figure 5) suggests that adsorbate hopping rates are a strong function of CC3 window size, but do not reveal the functional form of the relationship. To investigate this issue, we plotted a large number of Xe EA,C→V values calculated with the implicit ligand sampling TST method described above as a function of the corresponding CC3 window sizes (Figure 11).

Figure 11. EA for cage to void Xe hops at 300 K as a function of the CC3 windows size. The black curve shows the associated fit to an exponential function

Figure 11 shows a scatter plot of Xe cage to void hopping activation energies (EA,C→V) as a function of the CC3 window size. The relationship between EA and window size is well described by the exponential dependence suggested by Haldoupis et al.6 The associated fit to

26 ACS Paragon Plus Environment

Page 27 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

23 = J

KLM

, where dw is the window diameter and a and b are fitted parameters, is show in

Figure 11. This exponential dependence on the CC3 window sizes (which are normally distributed) explains the log-normal structure of the activation energy distributions collected during our ILS TST calculations (Figure 6). As a consistency check, we used the normal distribution fit parameters for CC3 window sizes at 300 K (Figure 2 inset) and the exponential fit parameters for the dependence of Xe EA,C→V values on window size (Figure 11) to reproduce the log-normal µ and σ parameters shown in Figure 4. This yields µ = 9.89 and σ = 0.533, which are within the uncertainty of the direct log-normal fit to the distribution in Xe EA,C→V observations from implicit ligand sampling at 300 K. Comparisons of adsorbate size to pore geometry can often predict the qualitative nature of diffusion through nanoporous materials. In CC3, the arene carbon window size is a reasonable approximation to the pore limiting diameter (PLD) found by Voronoi decomposition with Pore Blazer47 or Zeo++.14, 48 Adsorbates larger than the PLD generally diffuse slowly through the tightest constrictions in rigid nanoporous materials. For example, in our TST calculations in the rigid experimental CC3 structure, adsorbates with Lennard–Jones σ parameters14 (such as Rn, where σ = 4.17 Å) significantly larger than the experimental window size (3.62 Å) have very slow kC→V hopping rates. In more physically realistic fully flexible CC3 model, a single PLD value calculated from the experimental atomic coordinates does not describe the fluctuating nature of the window that restrict adsorbate diffusions. A more descriptive metric is the “pore limiting envelope”, which is the distribution in pore limiting diameters observed in flexible pores similar to the window size distribution. Holden et al. calculated the pore limiting envelope in the flexible CC3 structure with Zeo++.21 From this distribution, they calculated the percentage of flexible CC3 27 ACS Paragon Plus Environment

The Journal of Physical Chemistry

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 28 of 37

configurations with a transient PLD that exceeds the size of several light gases. Of the CC3 configurations sampled at 300 K, 58.7% have PLDs larger than the van der Waal diameter of Kr, while only 7.3% of configurations have PLDs which exceed the size of Xe. Using the pore limiting envelope analysis, the mechanism of adsorbate diffusion is transient pore connectivity which allows for opportunistic gas percolation. Holden et al. found that these transient pore connections are common for adsorbates comparable to the size experimental pore limiting diameter (such as Kr) but infrequent for larger adsorbates (such as Xe and Rn).19, 30 The pore limiting envelope provides a simple mechanism for the phenomena of “porosity without pores” observed for Xe and Rn diffusion in CC3.14, 30 However, the quantitative predictions of this approach are strongly dependent on what techniques are used to calculate PLDs and adsorbate radii. This model is not consistent with the treatment of Xe and Rn as soft LennardJones particles rather than hard spheres in MD simulations. Our results suggest a technique for interpreting our ILS TST calculations which explains opportunistic activated hopping of adsorbate molecules in CC3 without presuming the binary, “open–shut” mechanism characteristic of the pore limiting envelope construct. This analysis quantifies the importance of rare, highly favorable framework configurations to adsorbate diffusion. In our implicit ligand sampling TST calculations described above, the distributions in hopping activation energies EA,C→V and EA,V→C were fit to log-normal distributions as shown in Supplementary Tables S7-S16. From these distributions, the expectation values of the overall fully flexible hopping rates were calculated with Eq 6. In equation 6, the interval [0,∞] is integrated over to incorporate contributions from the entire distribution of activation energies to the overall hopping rate kC→V. If we instead integrate over the finite interval [0,EA], we find the partial expected value of the overall hopping rate from activation energies below EA. The 28 ACS Paragon Plus Environment

Page 29 of 37

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

The Journal of Physical Chemistry

quotient of the overall hopping rate kC→V (from Eq 6) and this partial expectation gives f, the fractional contribution of activation energies below some threshold EA to the overall rate: N 23 ) =



→